METHOD AND APPARATUS FOR REGISTRATION OF VIRTUAL ENDOSCOPIC IMAGES
CROSS-REFERENCE TO RELATED APPLICATIONS PRIORITY
Specific reference is hereby made to U.S. Provisional Application No. 60/495,304 (Attorney Docket No. 2003P12266US), entitled POINT SET MATCHING FOR PRONE- SUPINE REGISTRATION IN VIRTUAL ENDOSCOPY, filed August 14, 2003 in the name of Christophe Chefd'hotel and Bernhard Geiger, the inventors in the present application, and of which the benefit of priority is claimed and whereof the disclosure is hereby incorporated herein by reference. FIELD OF THE INVENTION
The present invention relates to the field of virtual endoscopy and, more particularly, to a method for automatically synchronizing views of volumetric images in virtual endoscopy, such as in virtual colonoscopy. BACKGROUND OF THE INVENTION
Virtual colonoscopy (VC) refers to a method of diagnosis based on computer simulation of standard, minimally invasive endoscopic procedures using patient specific three- dimensional (3D) anatomic data sets. Examples of current endoscopic procedures include bronchoscopy, sinusoscopy, upper gastro-intestinal endoscopy, colonoscopy, cystoscopy, cardioscopy, and urethroscopy. VC visualization of non-invasively obtained patient specific anatomic structures avoids risks, such as perforation, infection, hemorrhage, and so forth, associated with real endoscopy, and provides the endoscopist with important information prior to performing an actual endoscopic examination. Such information and understanding can minimize procedural difficulties, decrease patient morbidity, enhance training and foster a better understanding of therapeutic results.
In virtual endoscopy, 3D images are created from two-dimensional (2D) computerized tomography (CT) or magnetic resonance (MR) data, for example, by volume rendering. Present-day CT and MRI scanners typically produce a set of cross-sectional images which, in combination, produce a set of volume data. These 3D images are created to simulate images coming from an actual endoscope, such as a fiber optic endoscope.
In the field of virtual endoscopy and, more particularly, in the field of virtual colonoscopy, it is desirable to provide for synchronization of different endoscopic views,
such as views acquired in prone and supine positions of a patient. This facilitates the identification of features in the different views and facilitates, for example, the parallel study of prone and supine acquisitions in colon cancer screening.
Existing methods for synchronizing such views typically assume that a colon centerline is formed of a single connected component. See, for example, B. Acar, S. Napel, D.S. Paik, P. Li, J. Yee, C.F. Beaulieu, R.B. Jeffrey, Registration of supine and prone ct colonography data: Method and evaluation, Radiological Society of North America 87th Scientific Sessions, 2001; B. Acar, S. Napel, D.S. Paik, P. Li, J. Yee, R.B. Jeffrey, C.F. Beaulieu, Medial axis registration of supine and prone CT colonography data, Proceedings of 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2001), Istanbul, Turkey, 25-28 October 2001; and D. Nain, S. Haker, W.E.L. Grimson, E. Cosman Jr, W. Wells, H. Ji, R. Kikinis, C.-F. Westin, Intra-Patient Prone to Supine Colon Registration for Synchronized Virtual Colonoscopy, In Proceedings of MICCAI 2002, Tokyo, Japan, September 2002. BRDEF SUMMARY OF THE INVENTION
It is herein recognized that the assumption that a colon centerline is formed of a single connected component may not be an entirely valid assumption in various situations including, for example, the occurrence of partial occlusions which may result in cases of incomplete air insufflation and, for example, due to imperfect data segmentation.
It is an object of the present invention to provide a method facilitating parallel study of prone and supine acquisitions in applications such as colon cancer screening.
In accordance with an aspect of the present invention, a method is provided for automatically synchronizing endoscopic views of two volumetric images in virtual colonoscopy. In one aspect, the method utilizes optimal matching of two point sets, one for each volume of the respective volumetric images. These sets correspond to uniform samples of the colon centerline. Point correspondences are spatially extended to the entire data using a regularized radial basis function approximation.
In accordance with an aspect of the present invention, the method does not require an assumption that a colon centerline is formed of a single connected component. The method allows coping with partial occlusions, a situation often encountered in case of incomplete air insufflation or imperfect data segmentation.
In accordance with another embodiment, points are directly sampled on the colon surface.
In accordance with an embodiment of the present invention, two volumetric images are obtained using a colonoscopy protocol, including procedures such as bowel preparation and air insufflation. For each volume, a set of polygonal lines representing connected pieces of the colon centerline is made available. The distance from each point of the centerline to its closest point on the colon surface, otherwise herein also referred to as the colon radius, is considered to be known.
In accordance with an aspect of the present invention, a method for registration of, or between, virtual endoscopic images comprises deriving first and second volumetric images by an endoscopic protocol and representing the images by respective first and second volumetric image data sets; deriving respective centerline representations by connected line components; resampling the connected components to provide respective first and second sample sets; computing a descriptor for each sample; computing a similarity matrix using distances between the descriptors; and determining an optimal set of point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix.
In accordance with another aspect of the present invention, the step of deriving first and second volumetric images comprises a step of deriving the first and second volumetric images by a colonoscopy protocol.
In accordance with another aspect of the present invention, the step of deriving respective centerline representations comprises a step of deriving the respective centerline representations by connected polygonal line components; identifying a colon surface in each of the volumetric images; deriving respective colon radius information by determining the distance from each point of the centerline representation to a respectively closest point on the colon surface.
In accordance with another aspect of the present invention, a method comprises the steps of extrapolating the centerline point correspondences to a 3-dimensional/3-dimensional (3D/3D) transformation between the first and second volumetric images; and applying the 3D/3D transformation to transform a virtual endoscope position between the first and second volumetric images.
In accordance with another aspect of the present invention, a method for prone-supine registration of first and second volumetric images obtained by virtual colonoscopy and including respective centerline representations by connected polygonal line components and including respective colon radius information, comprises: resampling of the
connected components to provide respective first and second centerline sample sets; computing a descriptor for each sample; computing a similarity matrix using distances between the descriptors; determining an optimal set of centerline point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix; extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second volumetric images; and applying the 3D/3D transformation to transform a virtual endoscope position between the first and second volumetric images.
In accordance with another aspect of the present invention, a method for prone-supine registration of first and second volumetric images obtained by virtual colonoscopy represent by respective first and second volumetric datasets, including respective centerline representations by connected polygonal lines and including colon radius information, comprises resampling of the connected polygonal lines to provide respective first and second sample sets; computing a descriptor for each sample; computing a similarity matrix using distances between the descriptors; and determining an optimal set of point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix.
In accordance with another aspect of the present invention, method comprises the step of extrapolating the optimal set of centerline point correspondences to a 3D/3D transformation between the first and second volumetric images.
In accordance with another aspect of the present invention, the method comprises the step of applying the 3D/3D transformation to transform a virtual endoscope position betwςen the first and second volumetric images.
In accordance with another aspect of the present invention, a method for registration between virtual endoscopy images in first and second patient positions, comprises performing colon segmentation and feature extraction, including centerline and colon surface data for each of the images; resampling the centerline and colon surface data; computing respective local descriptors; pairing point correspondences on the centerlines between the first and second images by minimal cost matching; extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second images.
In accordance with another aspect of the present invention, a method for prone-supine registration of first and second volumetric images, obtained by virtual colonoscopy and
represented by first and second volumetric image data sets, including respective centerline representations by connected polygonal line components, and including respective colon radius data, comprises resampling of the connected line components to provide respective first and second sample sets; computing, for each sample, a descriptor comprising a vector of geometric features and an estimated value of the colon radius data; computing a similarity matrix using distances between the descriptors; applying a minimization procedure to the similarity matrix to determine an optimal set of correspondences between points of the first and second sample sets by applying an algorithm to the similarity matrix for minimizing the sum of distances between all corresponding points; extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second volumetric images; selecting a position for a virtual endoscope in one of the volumetric images; associating an orthogonal reference frame with the virtual endoscope; and applying the 3D/3D transformation to the orthogonal reference frame so as to derive a corresponding transformed reference frame for the virtual endoscope in the other of the volumetric images.
In accordance with another aspect of the present invention, a method for registration of virtual endoscopic images, the method comprises deriving first and second volumetric images by a colonoscopy protocol and representing the images by respective first and second volumetric image data sets; deriving respective centerline representations by connected polygonal line components; identifying a colon surface in each of the volumetric images; deriving respective colon radius information by determining the distance from each point of the centerline representation to a respectively closest point on the colon surface; resampling the connected components to provide respective first and second sample sets; computing a descriptor for each sample; computing a similarity matrix using distances between the descriptors; determining an optimal set of point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix.
In accordance with another aspect of the present invention, a method comprises extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second volumetric images.
In accordance with another aspect of the present invention, a method comprises applying the 3D/3D transformation to transform a virtual endoscope position between the first and second volumetric images.
In accordance with another aspect of the present invention, apparatus for prone-supine registration of first and second volumetric images obtained by virtual colonoscopy and including respective centerline representations by connected polygonal line components and including respective colon radius information, comprises apparatus for resampling of the connected components to provide respective first and second centerline sample sets; apparatus for computing a descriptor for each sample; apparatus for computing a similarity matrix using distances between the descriptors; apparatus for determining an optimal set of centerline point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix; apparatus for extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second volumetric images; and apparatus for applying the 3D/3D transformation to transform a virtual endoscope position between the first and second volumetric images.
In accordance with another aspect of the present invention, apparatus for registration of virtual endoscopic images, the apparatus comprises apparatus for deriving first and second volumetric images by an endoscopic protocol and representing the images by respective first and second volumetric image data sets; apparatus for deriving respective centerline representations by connected line components;apparatus for resampling the connected components to provide respective first and second sample sets; apparatus for computing a descriptor for each sample; apparatus for computing a similarity matrix using distances between the descriptors; and apparatus for determining an optimal set of point correspondences between the first and second sample sets by application of a minimization algorithm to the similarity matrix.
In accordance with another aspect of the present invention, a method for registration of virtual endoscopy images in first and second patient positions comprises performing colon segmentation and feature extraction, including centerline and colon surface data for each of the images; resampling the centerline and colon surface data; computing respective local descriptors; pairing point correspondences on the centerlines between the first and second images by minimal cost matching; extrapolating the centerline point correspondences to a 3D/3D transformation between the first and second images. The method also includes selecting a position for a virtual endoscope in one of the images; associating an orthogonal reference frame with the virtual endoscope; and applying the 3D/3D transformation to the orthogonal reference frame so as to derive a corresponding transformed reference frame for the virtual endoscope in the other of the images.
BRIEF DESCRD?TION OF THE DRAWINGS
These and other aspects of the invention will be more fully understood from the detailed description which follows, in conjunction with the drawings in which
Figure 1 shows a representation of a colon surface and its centerline, helpful to an understanding of the present invention;
Figure 2 shows a flow chart showing steps of an embodiment in accordance with the present invention;
Figure 3 shows uniform resampling of the centerline connected component in accordance with the present invention;
Figure 4 shows surface uniform sampling in accordance with the present invention; and
Figure 5 shows a block diagram of apparatus suitable for practicing the present invention. DETAILED DESCRD?TION OF THE INVENTION
Figure 1 shows a representation of the colon surface and its centerline. The centerline is indicated as a lighter colored thread within the image of the colon.
In outline, an embodiment of the method in accordance with the present invention comprises a first step of computation of local shape descriptors. The computation comprises a uniform resampling of all connected centerline components and, for each sample, a computation of a multi- valued descriptor, using local shape properties.
The method comprises a further step of optimal point set matching, comprising computing of a similarity matrix using distances between descriptors and determining of optimal point correspondences using a matching algorithm for weighted bipartite graphs (optimal assignment problem).
Further steps in accordance with the invention comprise performing registration and synchronization of the endoscopic views, a radial basis function approximation being used to extrapolate centerline correspondences to a 3D/3D transformation between volumes and, in the virtual endoscopy user interface, the endoscope' s position and orientation being updated using the resulting transformation.
Figure 2, shows in relation to first and second subject positions, for example, prone and supine data, the steps of colon segmentation and feature extraction 2 (Centerline / Surface); Centerline/surface resampling and computation of local descriptors 4; Point pairing using minimal cost matching 6; Extrapolation to 3D/3D transformation and
endoscope synchronization (position/orientation); and Synchronized (Prone/Supine) navigation and workflow. Figure 2 also shows a corresponding parallel graphical representation with juxtaposed prone and supine illustrations 2A - 8A.
In further detail, the step of data resampling comprises selecting a fixed number N of sampling points. For each dataset, the total length of its centerline components is computed, that is, added up, and divided by N-l. Each connected component is then resampled such that the arc-length between points corresponds to the previously computed value (Total Length / (N-l)). The position of new sample points is computed by linear interpolation.
Figure 3 shows uniform resampling of the centerline connected components: the original sampling is the solid line of Fig 3 (b) while the darkest line of Fig 3 (d) indicates the result after uniform resampling.
The step of descriptor computation is detailed next. A multi-valued descriptor is then computed for each sample point. This descriptor can be formed of the following attributes: a vector of scalar geometric features (curvature, torsion, distance to centroid) as well as the estimated distance to the colon surface (colon radius); and a list of vectors containing the Euclidean distances and orientations from the current point to all the other samples (of the same dataset). Orientations can be computed in a local (Frenet) frame or using a global coordinate system.
For explanatory material on Frenet formulas, curvature, torsion and related matters, see for example, a mathematical textbook such Chapter 15 of "Advanced Engineering Mathematics," second edition, by Michael D. Greenberg; Prentice-Hall, Upper Saddle River, New Jersey; 1998.
For the similarity matrix computation, anJVx N_ similarity matrix is built. It gives for each element of the first set a "distance" between its descriptor and the descriptors of all the elements of the second set. This distance between two descriptors is taken as the mean of distances between their corresponding attributes. It is noted that since distances between attributes may not have the same range, they are normalized before the mean is taken. Distances between scalar attributes are given by half of their squared differences and for vectors of Euclidean distances and orientations, the distance is evaluated using a statistical similarity measures, taking its opposite value and applying an offset if
necessary. In order to compute statistical similarity measures between two vectors their elements are assumed to be samples of two random variables X and Y.
In accordance with aspects of the present invention, three possible strategies are described next.
(a) The vectors are rank ordered and Spearman's rank correlation is computed. For a detailed description of this technique see, for example, the publication by W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Hannery, Numerical Recipes in C, Second Edition, Cambridge University Press, 1992. The Spearman rank correlation coefficient is defined by '≡ l- 6Y -2 (1)
where dl is the difference in rank of the vectors' z'-fh element. . The Spearman rank correlation coefficient provides a measure of the strength of the associations between two variables. For the Spearman rank correlation coefficient, see for example, CRC Concise Encyclopedia of Mathematics, Second Edition, Eric W. Weisstein; Chapman and Hall/CRC, New York, 2002; p. 2762 et seq.
(b) Histograms of the vector elements are computed and can be compared using the Kullback-Leibler divergence or the Chi-square measure. The Kullback-Leibler distance is defined as
where fx and fγ represent the probabilities (normalized histograms) of the corresponding variables X and Y, respectively.
For the Kullback-Leibler divergence or distance, see for example, Mathematics Handbook for Science and Engineering, Rade and Westergren, Studentlitteratur Birkhauser, Sweden, 1995; page 410.
The Chi-square measure is treated in mathematical texts; see for example, Applied Statistics for Engineers by William Volk, McGraw-Hill Book Company, Inc., New York, 1958; Chapter 5.
(c) The joint histogram of the two sets of vector elements is computed, and their similarity given by their Mutual Information. The Mutual Information is defined as
Here, fX (x,y) and fx (x), fγ (y) represent the joint and marginal probabilities of the pair of random variables (X,Y), respectively.
For the Mutual Information see, for example, the above-cited Mathematics Handbook for Science and Engineering, by Rade and Westergren, page 410 and the above-cited CRC Concise Encyclopedia of Mathematics by Weisstein.
With regard to bipartite graph matching, given the two sets of N points and the NxN similarity matrix, we try to find an optimal pairing (optimal assignment) which minimizes the sum of the distances between all corresponding points. This can be computed exactly using a weighted bipartite matching algorithm. A fast Augmenting Path technique can be applied to the similarity matrix previously obtained above, as described in the publication by R. Jonker, A. Volgenant, A Shortest Augmenting Path Algorithm for Dense and Sparse Linear Assignment Problems, Computing, 38:325-340, 1987.
Considering next the computation of the transformation, once a one-to-one correspondence is established between the two point sets, it is then used as a set of corresponding geometric landmarks. Landmark correspondences can be propagated to the entire space by computing two transformations (3D/3D mapping from the first volume to the second, and reciprocally) using a regularized radial basis function approximation. The regularization parameter is chosen empirically.
It is noted that the whole process: Descriptor computation, matching, computation of the transformation, can be iterated several times on updated version of the initial point sets.
The transformations can then be used to synchronize a prone and a supine view in the standard virtual colonoscopy workflow. The virtual endoscope is synchronized both in position and orientation using the following technique:
An infinitesimal orthogonal frame described by 4 points (one point for the origin and three for the basis vector extremities) is attached to the selected virtual endoscope. The transformation is applied to each point. The resulting frame, after orthogonalization, gives the corresponding endoscope position and orientation in the second dataset.
In another embodiment in accordance with the principles of the present invention, rather than using points on the centerline, one can also sample points uniformly on the surface of the colon. Figure 4 shows surface uniform sampling in accordance with this approach.
The descriptor can be updated accordingly to include surface specific features (such as the Gaussian and Mean curvature of the colon surface at this point). The rest of the registration procedure would remain the same. The centerline is not needed in this case.
The shape context framework for the realignment of 2D curves as discussed in the publication by S. Belongie, J. Malik, J. Puzicha, Shape Matching and Object Recognition Using Shape Context, IEEE Transactions on Pattern Analysis and Machine Intelligence, (24) 24:509-522, 2002 is of interest relative to an aspect of the present invention. However, it differs significantly for the ID matching approaches (warping based on dynamic programming, linear stretching/shrinking along the centerline path) previously used to perform intra-patient registration of prone and supine data in virtual colonoscopy. See also the two above-cited publications by Acar et al., Registration of supine and prone ct colonography data: Method and evaluation, and Medial axis registration of supine and prone CT colonography data; and the above-cited publication by Nain et al.
It will be understood that the present invention is intended to be practiced in conjunction with a programmable computer. Figure 5 shows a block diagram of apparatus suitable for practicing the present invention. Images are acquired by apparatus for image acquisition 50, as known in the art, in accordance with a protocol as earlier described. Such images, conveniently in digitized form, are stored and processed by a computer 52, in accordance with principles of the present invention. Processed images are viewable on image display apparatus 54, as known in the art, and may be further stored, processed, and/or transmitted by known communications techniques.
The invention has been described by way of exemplary embodiments. It will be apparent to one of ordinary skill in the art to which the present invention pertains that various changes and substitutions may be made without departing from the spirit of the invention.
These and like changes and substitutions are intended to be within the scope of the claims following.